Agent-FLAN: Enhancing Language Models for Agent Tasks
Introduction
Agent-FLAN is an innovative project focused on improving the capability of large language models (LLMs) to function effectively as agents. Although open-sourced LLMs have excelled in various natural language processing (NLP) tasks, their performance as agents lacks compared to API-based models. The challenge is to integrate agent abilities into general-purpose LLMs. This project identifies several key insights: the training data for agents is complex and shifts away significantly from the original pre-training data, LLMs learn agent tasks at different speeds, and current methods often introduce errors due to hallucinations. Agent-FLAN addresses these issues by refining the training process, allowing Llama2-7B, a model underpinning this project, to surpass existing methods by 3.5% in various evaluations. Additionally, Agent-FLAN reduces hallucinations and improves LLMs' agent capabilities while maintaining their general proficiency.
What's New
- March 21, 2024: The project paper was published on ArXiv.
- March 20, 2024: The dataset and model checkpoint for Agent-FLAN were released.
Agent-FLAN Series
The Agent-FLAN models are fine-tuned through the AgentInstruct and Toolbench datasets, using a specially designed data generation pipeline. This approach equips the models with robust skills for numerous agent tasks and tool operations.
Model and Dataset Availability
Agent-FLAN is available on platforms such as HuggingFace and OpenXLab. The training involves a mix of the AgentInstruct, ToolBench, and ShareGPT datasets, maintaining a conversational structure similar to Llama-2-chat models. The flexible and structured format enhances the efficiency of training LLMs for agent tasks.
Model Resources
- Agent-FLAN-7B: Can be accessed on HuggingFace and OpenXLab.
Dataset Resources
- Agent-FLAN Dataset: Available on HuggingFace.
Detailed Results
Agent-FLAN significantly outperforms previous approaches to agent tuning, as shown in evaluations of both known (held-in) and novel (held-out) tasks. The results are standardized against GPT-4 performance for a clearer comparison, marking a significant advancement in bridging the gap between open-source models and API-based models.
Acknowledgements
The success of the Agent-FLAN project relies on contributions from projects like Lagent and T-Eval, whose work underpins the Agent-FLAN development.
Citation
For scholarly use of Agent-FLAN, the formal citation is provided to credit the efforts of its developers and contributors, as published in the arXiv repository.
License
Agent-FLAN is made available under the Apache 2.0 license, promoting open access and further innovation in the field of advanced LLM agent functionality.